Unroll fixed-trip-count loops within mmt4d ukernel tile functions. (#16626)

mmt4d ukernels rely on fixed-trip-count for loops to do repetitive work
without encumbering source code. This relies on the compiler always
unrolling these for loops. We found in #16596 that that isn't always the
case.

The actual reason why unrolling failed to happen in the case in #16596
is actually that I forgot a `static inline` on the shared implementation
function for that ukernel; inlining it was necessary to reveal the
constancy of the trip count and thus the unrollability. So technically,
it might be enough to just add the missing `static inline`. But then,
when we forget it, that silently degrades performance as it did here. By
contrast, adding these `unroll(full)` pragmas causes a clang warning to
be generated when the loop can't be unrolled, which is how I realized
that I had forgotten the `static inline`. So as some layers of
defense-in-depth against suboptimal compilation of ukernel code, this PR
adds:
1. The missing `static inline` on those helper functions.
2. The `pragma unroll(full)` on all loops that should always be
unrolled.
3. `attribute(always_inline)` on top of `static inline` on those
functions that we really know should always be inlined and where failure
to inline would lead to failure to unroll.
11 files changed
tree: e87a6294dca490426f21cba5b13a2b7f8ad5aa27
  1. .devcontainer/
  2. .github/
  3. build_tools/
  4. compiler/
  5. docs/
  6. experimental/
  7. integrations/
  8. lib/
  9. llvm-external-projects/
  10. runtime/
  11. samples/
  12. tests/
  13. third_party/
  14. tools/
  15. .bazel_to_cmake.cfg.py
  16. .bazelignore
  17. .bazelrc
  18. .bazelversion
  19. .clang-format
  20. .dockerignore
  21. .git-blame-ignore-revs
  22. .gitignore
  23. .gitmodules
  24. .yamllint.yml
  25. AUTHORS
  26. BUILD.bazel
  27. CITATION.cff
  28. CMakeLists.txt
  29. configure_bazel.py
  30. CONTRIBUTING.md
  31. LICENSE
  32. README.md
  33. WORKSPACE
README.md

IREE: Intermediate Representation Execution Environment

IREE (Intermediate Representation Execution Environment, pronounced as “eerie”) is an MLIR-based end-to-end compiler and runtime that lowers Machine Learning (ML) models to a unified IR that scales up to meet the needs of the datacenter and down to satisfy the constraints and special considerations of mobile and edge deployments.

See our website for project details, user guides, and instructions on building from source.

CI Status IREE Discord Status

Project Status

IREE is still in its early phase. We have settled down on the overarching infrastructure and are actively improving various software components as well as project logistics. It is still quite far from ready for everyday use and is made available without any support at the moment. With that said, we welcome any kind of feedback on any communication channels!

Communication Channels

Related Project Channels

  • MLIR topic within LLVM Discourse: IREE is enabled by and heavily relies on MLIR. IREE sometimes is referred to in certain MLIR discussions. Useful if you are also interested in MLIR evolution.

Architecture Overview

IREE Architecture IREE Architecture

See our website for more information.

Presentations and Talks

License

IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.